Significant-bit precision of the ablated MRIQC IQMs

[oesteban]

Objective

Quantify the numerical precision of the ablated MRIQC-T1w IQMs with a significant-bit / significant-digit analysis derived from the MCA instrumentation (#1442), reporting how many bits (or digits) of each IQM are reproducible under floating-point perturbation, for the ablated pipeline versus the standard MRIQC.

Rationale

Bland–Altman (nipreps/fmriprep-equivalence#9) captures agreement between two pipelines; the significant-bit metric captures the intrinsic numerical precision of each pipeline's IQMs — how many bits survive floating-point perturbation. Together they give a complete numerical-equivalence picture: the ablated pipeline should not only agree with the reference on average but also carry comparable numerical precision.

Note

Precision and agreement are distinct properties: two pipelines can agree on average yet differ in how many bits of each IQM are reproducible. Significant bits is the deepest test of the numerical-equivalence oracle.

Methodological approach

1. From the MCA-instrumented runs (Instrument the ablated MRIQC for numerical variability #1442), compute the number of significant bits/digits per IQM for the ablated MRIQC-T1w and for the standard MRIQC.
2. Compare the ablated vs reference significant-bit distributions per IQM.
3. Report per-IQM significant-bit results, flagging any IQM where the ablated pipeline loses precision relative to the reference.

Outcomes/impact

1. A per-IQM significant-bit report showing whether the ablated MRIQC-T1w carries numerical precision comparable to the standard MRIQC.

Impact. The significant-bit dimension of the numerical-equivalence oracle, extending the verifier beyond agreement statistics — and, together with the Bland–Altman agreement analysis (nipreps/fmriprep-equivalence#9), the evidence that the ablated MRIQC is numerically equivalent to the reference before the methodology scales to the full fMRIPrep ablation.

Direct upstream issues

(issues this one depends on / builds upon)

• Instrument the ablated MRIQC for numerical variability #1442

Direct downstream issues

(issues that depend on this one)

None on the board yet — feeds the numerical-equivalence oracle informally.